📄 Sourcely - RAG-Powered PDF Research Assistant

An AI-powered research assistant that allows users to upload PDF documents and ask questions about their content. Sourcely uses Retrieval-Augmented Generation (RAG) to provide accurate answers with direct citations from the uploaded source.

📸 Preview

Watch the Sourcely Walkthrough

✨ Features

• 📂 smart parsing: extract text from pdf documents.
• 🔍 vector search: high-performance document retrieval using faiss.
• 💬 ai answers: answers questions using Groq's llama-3.3-70b-versatile model via context-only prompting.
• 📍 citations: every factual claim is cited with an exact [Chunk X] reference including page number.

🎯 How it works

---

The Sourcely Processing Pipeline

graph TD
    A[User Uploads PDF] --> B[Text Extraction]
    B --> C[Text Chunking]
    C --> D[Embedding Generation]
    D --> E[FAISS Index Building]
    E --> F[Ready for Queries]
    
    G[User Asks Question] --> H[Question Embedding]
    H --> I[Similarity Search in FAISS]
    I --> J[Retrieve Relevant Chunks]
    J --> K[LLM Response Generation]
    K --> L[Answer with Citations]
    
    subgraph "PDF Ingestion"
    B
    C
    D
    E
    end
    
    subgraph "Query Handling"
    H
    I
    J
    K
    end

Loading

🚀 Getting Started

The project is split into two main parts. Follow the instructions in each directory:

• Backend Setup - python, fastapi, mistral, and groq.
• Frontend Setup - react, vite, and tailwind.

💻 Tech Stack

• backend: fastapi, pdfplumber, faiss-cpu, mistral-embed, groq.
• frontend: react 19, vite, tailwind css, axios.
• llm: cloud inference — embeddings via mistral-embed, generation via Groq llama-3.3-70b-versatile.

💡 Troubleshooting

• PDF Extraction Failed: Ensure the PDF is not password-protected or purely image-based (OCR not currently supported).
• API Errors: Verify MISTRAL_API_KEY and GROQ_API_KEY are set correctly in backend/.env.
• Memory Issues: Large PDFs may require significant RAM for embedding generation.

---

📊 Benchmarks

Sourcely was evaluated across a diverse set of documents and queries to validate retrieval quality and response latency.

⚡ Retrieval Latency

Retrieval was benchmarked across two dimensions — isolated FAISS vector search, and full end-to-end query time including the Mistral embedding API call. Run via benchmark.py on a MacBook Air M1, 200 FAISS runs / 10 E2E runs per size, k=5.

FAISS-only (in-memory vector search, index.search isolated):

Document Size	Chunks	Mean	p95	p99
~25 pages	125	0.02 ms	0.01 ms	0.02 ms
~50 pages	250	0.02 ms	0.02 ms	0.03 ms
~100 pages	500	0.04 ms	0.05 ms	0.05 ms
150+ pages	750	0.06 ms	0.07 ms	0.09 ms

End-to-end (Mistral embed API call + FAISS search):

Document Size	Chunks	Mean	Median	p95
~25 pages	125	463 ms	394 ms	804 ms
~50 pages	250	450 ms	461 ms	555 ms
~100 pages	500	385 ms	377 ms	466 ms
150+ pages	750	371 ms	370 ms	461 ms

FAISS search itself is negligible (< 0.1 ms). The dominant cost is the Mistral embedding API network round-trip (~370–463 ms mean). E2E latency is under 400 ms median for documents ≥ 50 pages.

---

⚙️ Configuration Used

chunk_size  = 500   # characters per chunk
overlap     = 100   # character overlap between chunks
top_k       = 5     # chunks retrieved per query
embed_model = "mistral-embed"   # 1024-dim vectors
faiss_index = "IndexFlatL2"     # exact L2 search

These defaults were tuned to balance recall (retrieving the right context) against prompt length overhead sent to the LLM.

---

🧪 Methodology

1. Latency benchmark (backend/benchmark.py) — Two passes:
   • FAISS-only: 200 runs × 4 document scales using synthetic unit vectors (no API calls). Measures raw in-memory search speed.
   • End-to-end: 10 runs × 4 document scales using real Mistral mistral-embed API calls + FAISS search. Measures realistic query latency.
   • Timing via time.perf_counter() around get_embedding() + index.search().
2. Citation architecture — Context-only prompt enforces [Chunk X] citations and bans prior-knowledge use. Page numbers are returned directly from chunk metadata in every /query response.
3. Reproduce: cd backend && python benchmark.py

📝 License

MIT License - feel free to use and modify!

---

Note: This project is designed for local research and document analysis. Ensure you have the rights to the documents you upload.